Scholarlink

<p>Accurate prediction of the State of Health (SOH) is essential for the safety and reliability of lithium-ion batteries. Traditional single-model architectures encounter challenges related to noise resistance and prediction accuracy when processing complex battery aging patterns. This paper proposes a DSwin-Transformer architecture that integrates relaxation voltage analysis with deep learning techniques. The method combines a denoising autoencoder (DAE) for feature dimensionality reduction with a hierarchical window attention mechanism to capture local degradation details and global aging dependencies. In this framework, the relaxation voltage is selected as the primary aging feature. Additionally, representative auxiliary aging features are identified from variables such as constant current charging time and voltage area using Pearson correlation analysis, mutual information, and SHAP values. Experimental validation across the CALCE and Tongji datasets demonstrates performance with mean absolute errors below 0.95%, showing improvements over Convolutional Neural Networks (CNN)-Gated Recurrent Units, CNN-Long Short-Term Memory, and CNN-Transformer baseline&#xa0;methods across various battery cells and training volumes. Ablation studies reveal the contribution of the DAE module, with the root mean square error (RMSE) improving from 0.179 to 0.0084 on the CALCE dataset for complex degradation patterns. The model maintains accuracy during both capacity decay and recovery periods while requiring a footprint of 15.6&#xa0;MB. The results indicate that combining relaxation voltage features with adapted computer vision architectures provides a practical approach for predicting battery SOH.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A DSwin-transformer-based SOH prediction method for lithium-ion batteries using relaxation voltages

  • 期刊论文

  • 发表日期：

    2025年9月19日

查看全文
  • Simin Yang,
  • Xiaojun Tan,
  • Jiagen Li,
  • Yuqian Fan,
  • Ziyu Zhao,
  • Binbin Chen,
  • Quanxue Guan

摘要

Accurate prediction of the State of Health (SOH) is essential for the safety and reliability of lithium-ion batteries. Traditional single-model architectures encounter challenges related to noise resistance and prediction accuracy when processing complex battery aging patterns. This paper proposes a DSwin-Transformer architecture that integrates relaxation voltage analysis with deep learning techniques. The method combines a denoising autoencoder (DAE) for feature dimensionality reduction with a hierarchical window attention mechanism to capture local degradation details and global aging dependencies. In this framework, the relaxation voltage is selected as the primary aging feature. Additionally, representative auxiliary aging features are identified from variables such as constant current charging time and voltage area using Pearson correlation analysis, mutual information, and SHAP values. Experimental validation across the CALCE and Tongji datasets demonstrates performance with mean absolute errors below 0.95%, showing improvements over Convolutional Neural Networks (CNN)-Gated Recurrent Units, CNN-Long Short-Term Memory, and CNN-Transformer baseline methods across various battery cells and training volumes. Ablation studies reveal the contribution of the DAE module, with the root mean square error (RMSE) improving from 0.179 to 0.0084 on the CALCE dataset for complex degradation patterns. The model maintains accuracy during both capacity decay and recovery periods while requiring a footprint of 15.6 MB. The results indicate that combining relaxation voltage features with adapted computer vision architectures provides a practical approach for predicting battery SOH.